self.md radar — 2026-08-26
Apple has put two very different local-AI machines on the same desk. One is a small Mac mini with a 32GB ceiling. The other is a Mac Studio that can hold 512GB in unified memory and move it at 1.2TB/s. That is not a spec-sheet footnote: it draws a much clearer line between running a helpful model nearby and treating a Mac as a serious private compute box.
1. the Mac Studio has entered the local-model argument
Apple’s M5 Ultra announcement puts an unusually large number on the table: up to 512GB of unified memory and 1.2TB/s of memory bandwidth in the new Mac Studio. Apple says the quad-die chip can keep models with hundreds of billions of parameters entirely in local memory; the Mac Studio page is where the configuration stops being a keynote adjective and becomes a machine somebody can actually price.
The interesting change is not that a desktop got faster. Local inference has usually meant choosing a smaller model, aggressive quantization, or a fiddly multi-machine setup. A 512GB unified pool changes the size of the things that can plausibly stay on one person’s desk. It also changes the constraint. The question moves from “can I install a local model?” to whether the model, its context, files, indexes, and runtime can fit together without quietly falling back to somebody else’s server.
That does not make the Studio a democratic little home server. It will be expensive, power-hungry, and still welded to Apple’s hardware choices. But it is a material shift in the equipment available to a person who wants a private working model to be more than an emergency offline toy.
reading: Apple’s M5 Ultra release · Mac Studio configurations
2. the Mac mini is the smaller, more honest boundary
The same release gives the new M6 Mac mini up to 32GB of unified memory, 170GB/s of bandwidth, a 12-core GPU with Neural Accelerators, and a dual 16-core Neural Engine. Apple’s Mac mini page frames it as a general desktop; the press release explicitly points to code compilation, indexing, simulators, and on-device model work.
Thirty-two gigabytes is enough to make local assistance ordinary for a lot of tasks. It is not enough to blur the difference between a compact helper and a large private runtime. That limit is useful. It makes the handoff visible: a small model can sit beside the files; a bigger one still wants either a larger machine or a remote bill.
The ownership question starts there. “Runs on my Mac” should mean more than a menu item. It should name the model size, the data that leaves the device, and the moment the workflow crosses out of the room.
reading: Apple’s M6 release · Mac mini specifications
3. memory bandwidth has become part of the agent brief
Apple’s press release gives the M6 up to 170GB/s of unified-memory bandwidth and the M5 Ultra 1.2TB/s. Those figures are not a clean prediction of model speed: model architecture, quantization, software, context length, and the runtime all interfere. But they are now part of the practical brief, alongside RAM capacity and price.
That is a healthier conversation than treating any neural-engine number as a verdict. A local agent that searches files, carries a large context, runs tools, and leaves a review trail is a system, not a single benchmark. The slow failure is buying the machine for a dazzling model demo, then discovering the actual stack does not fit, does not run well, or cannot be inspected when it makes a mess.
Apple has made the physical side of that stack much more capable. It has not made the design problem go away. The useful purchase question is still boring: what must stay local, what must remain recoverable, and what evidence will tell you that the machine did the thing it claims?
reading: Apple’s silicon details · Mac Studio · Mac mini
more to read
- A Go dependency wrote
AGENTS.mdmid-build and got Codex to hide the change — a concrete reminder that instruction files are not a security boundary when dependencies can write them after review. - LLM Agents Perform Controlled Experiments Using Simulation Models — a new paper that puts the agent beside a simulator and compares interventions instead of asking it to sound plausible about a process.
- AI Agents Push Humans Out of the Loop — a position paper worth reading for its narrow question: what does a person need in order to keep judging an automated system, rather than merely approving it?
left on the table
- RENDER makes a sharp point about how memory is presented to a model, but it was submitted in June and is not today’s change.
- OpenBiliClaw is surging again, yet it was already in the Radar recently and its current momentum is not new evidence of a changed product or practice.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Best Claude Code plugins — choose the Claude-specific extensions worth installing, and the ones to skip